A Pixel-Level Classification Method and System for Rasterized Point Clouds Based on Deep Learning

Abstract

The present invention relates to a rasterized point cloud pixel-level classification method and system based on deep learning, including splitting two-dimensional raster data into blocks of different sizes, finding the best initial receptive field as data input, and extracting the features of the segmented LIDAR two-dimensional raster data through a network model using a first model as the backbone network, and training the model; selecting the information features extracted by the backbone network through an information refinement block, cross-fusing different-level feature maps, and expanding the receptive field of the shallow feature maps; aggregating the features of related points within the neighborhood, strengthening the implicit space feature representation of each pixel point, and thereby improving the classification accuracy of LIDAR data. The present invention initializes and trains the model parameters through transfer learning, introduces an information refinement and cross-fusion module to select the information features extracted by the backbone network, and uses object context representation to complete feature enhancement, ultimately improving the classification effect of the rasterized lidar data.

Description

Technical Field

[0001] The present invention relates to the technical field of radar data processing, and in particular, to a rasterized point cloud pixel-level classification method and system based on deep learning. Background Art

[0002] Due to its unique advantages, Light Detection and Ranging (LiDAR) has a wide range of civilian and military applications and a bright future. It can be used independently or in conjunction with optical data sources such as aerial images. However, in contrast to raw images for which many effective tools are available, they are difficult to process. To overcome the special nature of the unstructured and sparse nature of LiDAR point cloud data, the current standard method is to convert the point cloud into a digital elevation model (DEM), represented as a two-dimensional raster. Such a raster can be used analogously to an optical image, for example, in conjunction with a 2-D convolutional neural network (CNN) for semantic segmentation. In many studies, land cover maps are often simplified to digital elevation models and used as additional information for fusion with multi-spectral or hyperspectral images.

[0003] With the latest developments in sensing technology, multi-modal data has become increasingly easily available for various fields, especially in Remote Sensing, where there are many data types such as multi-spectral imaging (MSI), hyperspectral imaging (HSI), Light Detection and Ranging (LiDAR), etc. To make full use of hyperspectral images (HSI) and LiDAR data, many different methods and strategies have been proposed over the past decade. Currently, many classification methods using HSI data have been developed. Since different source data have specific advantages, various classification fusion strategies have also been proposed to combine multiple features of different data sources. However, little consideration has been given to how to achieve better results in the classification method of rasterized LIDAR data. Therefore, how to accurately segment and classify LiDAR point clouds is an issue that needs to be considered at this stage.

[0004] It should be noted that the information disclosed in the above Background Art section is only used to enhance the understanding of the background of the present disclosure, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a rasterized point cloud pixel-level classification method and system based on deep learning, which solves the deficiencies existing in the prior art.

[0006] The purpose of the present invention is achieved by the following technical solutions: A rasterized point cloud pixel-level classification method based on deep learning, the classification method comprising:

[0007] Feature extraction step: The two-dimensional raster data is sliced into patches of different sizes, the best initial receptive field of the rasterized LIDAR data is found as the data input, and the features of the sliced LIDAR two-dimensional raster data are extracted through a network model using the first model as the backbone network, and the model is trained;

[0008] Feature refinement step: Select the LIDAR information features extracted by the backbone network through the information refinement block, cross-fuse the feature maps of different levels, and expand the receptive field of the shallow LIDAR feature maps;

[0009] Feature enhancement step: Aggregate the features of related points in the neighborhood and strengthen the feature representation of each pixel point in the latent space;

[0010] Finally, by extracting, refining, fusing, and enhancing the features of the rasterized LIDAR data, the most effective features are selected from the original features to train the model, improving the classification accuracy of the model for LIDAR data, thereby enhancing the classification and fusion of HSI and LIDAR data.

[0011] The specific steps of the feature extraction step include:

[0012] The two-dimensional raster data is sliced into data patch blocks with a radius of r centered on pixels as input data, and the data patches are sent into a 2-DCNN tunnel. Spatial features are extracted from the surrounding area of the central target pixel through the Resnet50 model; the Resnet50 model includes two basic blocks, namely Conv Block and Identity Block. Multiple Conv Block basic blocks are connected in parallel to change the dimension of the network, and multiple Identity Block basic blocks are connected in series to deepen the depth of the network;

[0013] The Resnet50 model contains 4 convolutional blocks, and each convolutional block has 3, 4, 6, and 3 bottleneck structures respectively. The spatial resolution of the feature map is reduced through the convolutional block and the max pooling layer. Finally, the feature maps (F3 and F4) output by the second and third convolutional blocks are selected as the input for the feature refinement step.

[0014] The feature refinement step includes:

[0015] The feature maps F output by the second and third convolutional blocks in the feature extraction step in are used as the input of the information refinement block, and a 3×3 convolutional layer with BN and ReLU is used to reduce the number of channels of the input feature map F in so as to reduce the computational amount of the network;

[0016] Unify the number of channels of the feature maps from different stages for further fusion. Through global average pooling operation, and use a 3×3 convolutional layer with BN and Sigmoid activation functions to generate a global weight vector Mc, and multiply the global weight vector Mc with the feature map F in to generate the output of the channel attention refinement block, and capture more accurate semantic information through the channel attention refinement block;

[0017] The feature maps (F3 and F4) output by the second and third convolutional blocks are refined to obtain a low-level feature map F low and a high-level feature map F high . Input the two feature maps into the cross-fusion block, and use bilinear interpolation operation to change the resolution of the high-level feature map F high to the same resolution as the low-level feature map F low , and refine and upsample the high-level feature map through a 3×3 convolutional layer with BN and ReLU;

[0018] Through concatenation operation, concatenate the upsampled high-level feature map F high and the low-level feature map F low , and generate the output of the cross-fusion block through a 1×1 convolutional layer with BN and ReLU.

[0019] The feature enhancement step includes:

[0020] After information refinement and fusion, divide the obtained feature map into k object regions {M1, M2, …, M k}, aggregate all the pixels in each object region, and weight them according to their degree of belonging to the object region to obtain the object region feature f k ;

[0021] Calculate the relationship ω ik between each pixel and each object region. Through aggregating the representations of k object regions and through the object region feature f k and the relationship ω ik between the pixel and the region, implement feature enhancement through a 1×1 convolutional layer with BN and ReLU, and finally obtain the features of the original representation and the object context representation after feature enhancement.

[0022] A rasterized point cloud pixel-level classification system based on deep learning, which includes a feature extraction unit, a feature refinement unit, and a feature enhancement unit;

[0023] The feature extraction unit: It is used to divide the two-dimensional raster data into patches of different sizes, find the best initial receptive field of the rasterized point cloud as the data input, and extract the features of the segmented LIDAR two-dimensional raster data through a network model with the first model as the backbone network, and train the model;

[0024] The feature refinement unit: It is used to select the information features extracted by the backbone network through the information refinement block, cross-fuse different-level feature maps, and expand the receptive field of the shallow feature maps;

[0025] The feature enhancement unit: It is used to aggregate the features of related points in the neighborhood, strengthen the hidden space feature representation of each pixel point, and improve the classification accuracy of the LIDAR data to enhance the classification and fusion of HIS and LIDAR data.

[0026] The feature refinement unit includes an information refinement subunit and a cross-fusion subunit;

[0027] The information refinement subunit: It is used to use the feature map F output by the second and third convolutional blocks in the feature extraction unit in as the input of the information refinement block, and use a 3×3 convolutional layer with BN and ReLU to reduce the number of channels of the input feature map F in to reduce the computational amount of the network; unify the number of channels of the feature maps from different stages for further fusion, through global average pooling operation, and use a 3×3 convolutional layer with BN and Sigmoid activation function to generate a global weight vector Mc, and multiply the global weight vector Mc and the feature map F in to generate the output of the channel attention refinement block, and capture more accurate semantic information through the channel attention refinement block;

[0028] The cross-fusion subunit: It is used to input the feature maps (F3 and F4) output by the second and third convolutional blocks, and after information refinement, obtain the low-level feature map F low and the high-level feature map F high into the cross-fusion block, use bilinear interpolation operation to change the resolution of the high-level feature map F high to the same resolution as the low-level feature map F low and refine the upsampled high-level feature map through a 3×3 convolutional layer with BN and ReLU; concatenate the upsampled high-level feature map F high and the low-level feature map F low through concatenation operation, and generate the output of the cross-fusion block through a 1×1 convolutional layer with BN and ReLU.

[0029] The feature enhancement unit includes an object region feature generation subunit and a calculation aggregation subunit;

[0030] The object region feature generation subunit: It is used to divide the feature map obtained after information refinement and fusion into k object regions {M1, M2, …, M k}, aggregate all the pixels of each object region, and weight them according to the degree of their belonging to the object region, to obtain the object region feature f k ;

[0031] The calculation and aggregation subunit: It is used to calculate the relationship ω between each pixel and each object region ik , and through aggregating the representations of k object regions and through the object region feature f k and the relationship ω between the pixel and the region ik , and through a 1×1 convolutional layer with BN and ReLU to achieve feature enhancement, and finally obtain the features of the original representation and the object context representation after feature enhancement.

[0032] The present invention has the following advantages: A rasterized point cloud pixel-level classification method and system based on deep learning, initializes the training of model parameters through transfer learning, introduces an information refinement and cross-fusion module to select the information features extracted by the backbone network, and uses the object context representation to complete feature enhancement, and finally improves the classification effect of the rasterized lidar data. Description of the Drawings

[0033] Figure 1 is a schematic flow chart of the method of the present invention;

[0034] Figure 2 is a schematic diagram of the feature extraction structure;

[0035] Figure 3 is a schematic diagram of the information refinement structure;

[0036] Figure 4 is a schematic diagram of the information fusion result;

[0037] Figure 5 is a schematic diagram of the feature enhancement structure. Detailed Embodiment

[0038] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this application. Apparently, the described embodiments are only a part rather than all of the embodiments of this application. Components of the embodiments of this application usually described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application provided below with reference to the accompanying drawings is not intended to limit the protection scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative efforts fall within the protection scope of this application. The following further describes the present invention with reference to the accompanying drawings.

[0039] As Figure 1 shown, one implementation of the present invention relates to a rasterized point cloud pixel-level classification method based on deep learning, which includes:

[0040] Feature extraction step: The two-dimensional raster data is sliced into patches of different sizes, the best initial receptive field of the rasterized LIDAR data is found as the data input, and the features of the sliced LIDAR two-dimensional raster data are extracted through a network model using the first model as the backbone network, and the model is trained.

[0041] Feature refinement step: The information refinement block is used to select the LIDAR information features extracted by the backbone network, and different-level feature maps are cross-fused to expand the receptive field of the shallow LIDAR feature maps.

[0042] Feature enhancement step: Aggregate the features of related points in the neighborhood to enhance the hidden space feature representation of each pixel point.

[0043] Finally, by extracting, refining, fusing, and enhancing the features of the rasterized LIDAR data, the most effective features are selected from the original features for model training, improving the classification accuracy of the model for LIDAR data, and thus enhancing the classification and fusion of HSI and LIDAR data.

[0044] Furthermore, since the dataset after rasterization of the original sparse point cloud is adopted, it is no longer restricted by the unstructured and large amount of data of the original point cloud. Therefore, it can be analogized to optical image processing. First, these two-dimensional gratings are cut into patches of different sizes for experiments to find the optimal size; part of the model and weights of the classic network ResNet50 are introduced and used as the backbone network for point cloud feature extraction for transfer learning. In this process, using the idea of transfer learning, the weights of the pre-trained ResNet50 are adopted, and the features obtained by training with large-scale data are used to initialize the model parameters, so as to accelerate the convergence speed of the lidar point cloud data classification model and improve the generalization ability of the classification model.

[0045] The feature extraction steps specifically include:

[0046] Taking the pixel as the center, the two-dimensional raster data is cut into data patch blocks with a radius r (such as 4) as the input data, and the data patch (size ksize = 2×r + 1) is sent into the 2-D CNN tunnel, and spatial features are extracted from the surrounding domain of the central target pixel through the ResNet50 model; as Figure 2 shown, the ResNet50 model includes two basic blocks, namely Conv Block and Identity Block. Multiple Conv Block basic blocks are connected in parallel to change the dimension of the network, and multiple Identity Block basic blocks are connected in series to deepen the depth of the network; there are 4 convolutional blocks in the ResNet50 model, and each convolutional block has 3, 4, 6, and 3 bottleneck structures respectively. The spatial resolution of the feature map is reduced through the convolutional block and the max pooling layer, and then four convolutional blocks are used to encode more accurate semantic information.

[0047] The input of ResNet50 usually has three channels. The lidar data used in the present invention has 1 and 2 channels respectively. The outputs (F3 and F4) of the 2nd and 3rd Blocks (convolutional blocks) are selected as the input of the next-stage information refinement module.

[0048] Furthermore, after passing through the backbone network, 2D raster features of point clouds at different levels are extracted. An information refinement block is used to select information, enabling the network to capture more accurate semantic information of point clouds at various scales; and the feature maps at different levels are cross-fused to expand the receptive field of the 2D raster feature map of the shallow point cloud. In other words, the low-level 2D raster feature map of the point cloud will have a receptive ability similar to that of the high-level 2D raster feature map of the point cloud, and the high-level feature map is used to compensate for the receptive field of the low-level feature map. In addition, the cross-fusion block has a very simple structure and brings little computational burden, thus having the effective characteristic of capturing more accurate small-scale 2D raster semantic information of point clouds; in this process, based on the channel attention mechanism, an information refinement block is first designed to select the information features of the rasterized point cloud by introducing global weights to different channels of the 2D raster feature map of the point cloud. Then, its semantic information capture ability is significantly improved through the cross-fusion block.

[0049] The feature refinement steps include:

[0050] Taking the feature maps F output by the second and third convolutional blocks in the feature extraction step in as the input of the information refinement block, as Figure 3 shown, using a 3×3 convolutional layer (Convolution) with BN (batch normalization layer) and ReLU (rectified linear layer) to reduce the number of channels of the input feature map F in to reduce the computational amount of the network;

[0051] Unifying the number of channels of the feature maps from different stages for further fusion, through global average pooling operation, and using a 3×3 convolutional layer with BN and Sigmoid activation function to generate a global weight vector Mc, multiplying the global weight vector Mc and the feature map F in after passing through the convolutional layer to generate the output of the channel attention refinement block, and capturing more accurate semantic information through the channel attention refinement block;

[0052] The feature maps (F3 and F4) output by the second and third convolutional blocks, after information refinement, obtain a low-level feature map F low and a high-level feature map F high , inputting the two feature maps into the cross-fusion block, as Figure 4 shown, using bilinear interpolation operation to change the resolution of the high-level feature map F high to the same resolution as the low-level feature map F low , refining and upsampling the high-level feature map through a 3×3 convolutional layer with BN and ReLU; concatenating the upsampled high-level feature map F high and the low-level feature map F low through concatenation operation, and generating the output of the cross-fusion block through a 1×1 convolutional layer with BN and ReLU.

[0053] Furthermore, the channel attention mechanism is utilized to find out the context features of the corresponding objects of the two-dimensional raster pixel points of the point cloud and used for feature enhancement; finally, the classification accuracy of the Lidar branch is improved to enhance the effect of the classification and fusion of His and Lidar data. In this process, by aggregating the features of relevant points in the neighborhood and considering the relationship between the current two-dimensional raster pixel point of the point cloud and the object area to which it belongs, the feature representation of the point cloud data is strengthened.

[0054] As Figure 5 shown, the feature enhancement steps include:

[0055] After information refinement and fusion, the obtained feature map is divided into k object areas {M1, M2, …, M k}, each object area M k corresponds to the kth category, and the obtained result is regarded as a rough semantic segmentation result. All the pixels in each object area are aggregated and weighted according to the degree to which they belong to the object area to obtain the object area feature f k ;

[0056] Calculate the relationship ω ik between each pixel and each object area, that is, the softmax weight. By aggregating the representation of the kth object area and considering the relationship between each pixel point and the object area to which it belongs to enhance its representation, that is, through the object area feature f k and the relationship ω ik between the pixel and the area, the feature enhancement is realized through a 1×1 convolutional layer with BN and ReLU, and finally the features of the original representation and the object context representation after feature enhancement are obtained.

[0057] Another embodiment of the present invention relates to a rasterized point cloud pixel-level classification system based on deep learning, which includes a feature extraction unit, a feature refinement unit, and a feature enhancement unit;

[0058] The feature extraction unit: is used to cut the two-dimensional raster data into patches of different sizes, find the best initial receptive field of the rasterized point cloud as the data input, and extract the features of the segmented LIDAR two-dimensional raster data through a network model using the first model as the backbone network, and train the model;

[0059] The feature refinement unit: is used to select the information features extracted by the backbone network through the information refinement block, cross-fuse different-level feature maps, and expand the receptive field of the shallow feature maps;

[0060] The feature enhancement unit: It is used to aggregate the features of relevant points in the neighborhood, enhance the feature representation of each pixel point in the latent space, and improve the classification accuracy of the HIS and LIDAR data classification fusion by enhancing the classification accuracy of the LIDAR data.

[0061] The feature refinement unit includes an information refinement subunit and a cross-fusion subunit;

[0062] The information refinement subunit: It is used to take the feature maps F output by the second and third convolutional blocks in the feature extraction unit in as the input of the information refinement block, and use a 3×3 convolutional layer with BN and ReLU to reduce the number of channels of the input feature map F in to reduce the computational amount of the network; unify the number of channels of the feature maps from different stages for further fusion, through global average pooling operation, and use a 3×3 convolutional layer with BN and Sigmoid activation function to generate a global weight vector Mc, and multiply the global weight vector Mc and the feature map F in to generate the output of the channel attention refinement block, and capture more accurate semantic information through the channel attention refinement block;

[0063] The cross-fusion subunit: It is used to take the feature maps (F3 and F4) output by the second and third convolutional blocks, and after information refinement, obtain the low-level feature map F low and the high-level feature map F high and input them into the cross-fusion block, use bilinear interpolation operation to change the resolution of the high-level feature map F high to the same resolution as the low-level feature map F low and refine the upsampled high-level feature map through a 3×3 convolutional layer with BN and ReLU; cascade the upsampled high-level feature map F high and the low-level feature map F low and generate the output of the cross-fusion block through a 1×1 convolutional layer with BN and ReLU.

[0064] The feature enhancement unit includes an object region feature generation subunit and a computational aggregation subunit;

[0065] The object region feature generation subunit: It is used to divide the feature map obtained through information refinement and fusion into k object regions {M1, M2,..., M k}, aggregate all the pixels in each object region, and weight them according to the degree to which they belong to the object region to obtain the object region feature f k ;

[0066] The computational aggregation subunit: It is used to calculate the relationship ω ik between each pixel and each object region, and through aggregating the representations of k object regions and through the object region feature fk and the relationship ω between pixels and regions ik , and the enhancement of features is achieved through a 1×1 convolutional layer with BN and ReLU, and finally the features of the original representation and the object context representation with enhanced features are obtained.

[0067] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be changed within the scope of the concept described herein through the above teachings or the techniques or knowledge in related fields. And the changes and alterations made by those skilled in the art that do not depart from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A rasterized point cloud pixel-level classification method based on deep learning, characterized in that: The classification method includes: Feature extraction step: The two-dimensional raster data is sliced into patches of different sizes, the best initial receptive field of the rasterized LIDAR data is found as the data input, and the features of the sliced LIDAR two-dimensional raster data are extracted through a network model using the first model as the backbone network, and the model is trained. Feature refinement step: The information refinement block is used to select the LIDAR information features extracted by the backbone network, and different-level feature maps are cross-fused to expand the receptive field of the shallow LIDAR feature maps. Feature enhancement step: Aggregate the features of related points in the neighborhood to enhance the implicit space feature representation of each pixel point. Finally, by extracting, refining, fusing, and enhancing the features of the rasterized LIDAR data, the most effective features are selected from the original features to train the model, improving the classification accuracy of the model for LIDAR data, thereby enhancing the classification and fusion of HSI and LIDAR data.

2. The rasterized point cloud pixel-level classification method based on deep learning according to claim 1, characterized in that: The specific steps of the feature extraction step include: The two-dimensional raster data is sliced into data patch blocks with a radius of r centered on pixels as the input data, and the data patches are sent into the 2-DCNN tunnel, and the spatial features are extracted from the surrounding area of the central target pixel through the Resnet50 model; the Resnet50 model includes two basic blocks, namely Conv Block and Identity Block. Multiple Conv Block basic blocks are connected in parallel to change the dimension of the network, and multiple Identity Block basic blocks are connected in series to deepen the depth of the network. The Resnet50 model contains 4 convolutional blocks, and each convolutional block has 3, 4, 6, and 3 bottleneck structures respectively. The spatial resolution of the feature map is reduced through the convolutional block and the max pooling layer. Finally, the feature maps (F3 and F4) output by the second and third convolutional blocks are selected as the input of the feature refinement step.

3. A pixel-level classification method for rasterized point clouds based on deep learning according to claim 1, characterized in that: The feature refinement step includes: Take the feature maps F output by the second and third convolutional blocks in the feature extraction step in as the input of the information refinement block, and use a 3×3 convolutional layer with BN and ReLU to reduce the input feature map F in in terms of the number of channels, thereby reducing the computational complexity of the network; Unify the number of channels of the feature maps from different stages for further fusion. Through global average pooling operation, and use a 3×3 convolutional layer with BN and Sigmoid activation functions to generate a global weight vector Mc. Multiply the global weight vector Mc with the feature map F in to generate the output of the channel attention refinement block, and capture more accurate semantic information through the channel attention refinement block; The feature maps (F3 and F4) output by the second and third convolutional blocks are refined to obtain the low-level feature map F low and the high-level feature map F high . The two feature maps are input into the cross-fusion block, and the bilinear interpolation operation is used to change the resolution of the high-level feature map F high to the same resolution as the low-level feature map F low , and the upsampled high-level feature map is refined through a 3×3 convolutional layer with BN and ReLU; The upsampled high-level feature map F is cascaded through operations high and the low-level feature map F low are cascaded, and the output of the cross-fusion block is generated through a 1×1 convolutional layer with BN and ReLU 4. A rasterized point cloud pixel-level classification method based on deep learning according to claim 1, characterized in that: The feature enhancement step includes: After information refinement and fusion, the obtained feature map is divided into k object regions {M1, M2, …, M k}, and all pixels in each object region are aggregated and weighted according to the degree of their belonging to the object region to obtain the object region feature f k ; Calculate the relationship ω between each pixel and each object region ik , by aggregating the representations of k object regions and through the object region feature f k and the relationship ω between the pixel and the region ik , the enhancement of the feature is realized through a 1×1 convolutional layer with BN and ReLU, and finally the features of the original representation and the object context representation with enhanced features are obtained.

5. A rasterized point cloud pixel-level classification system based on deep learning, characterized in that: It includes a feature extraction unit, a feature refinement unit, and a feature enhancement unit. The feature extraction unit: is used to slice the two-dimensional raster data into patches of different sizes, find the best initial receptive field of the rasterized point cloud as the data input, and extract the features of the sliced LIDAR two-dimensional raster data through a network model using the first model as the backbone network, and train the model. The feature refinement unit: is used to select the information features extracted by the backbone network through the information refinement block, cross-fuse different-level feature maps, and expand the receptive field of the shallow feature maps. The feature enhancement unit: is used to aggregate the features of related points in the neighborhood, enhance the implicit space feature representation of each pixel point, and enhance the classification and fusion of HIS and LIDAR data by improving the classification accuracy of LIDAR data.

6. The rasterized point cloud pixel-level classification system based on deep learning according to claim 5, wherein: The feature refinement unit includes an information refinement subunit and a cross-fusion subunit. The information refinement subunit: is used to take the feature maps F output by the second and third convolutional blocks in the feature extraction unit in as the input of the information refinement block, and uses a 3×3 convolutional layer with BN and ReLU to reduce the input feature map F in in terms of the number of channels so as to reduce the computational amount of the network; unify the number of channels of the feature maps from different stages for further fusion, through global average pooling operation, and use a 3×3 convolutional layer with BN and Sigmoid activation function to generate a global weight vector Mc, and multiply the global weight vector Mc and the feature map F in to generate the output of the channel attention refinement block, and capture more accurate semantic information through the channel attention refinement block; The cross-fusion subunit: is used to refine the feature maps (F3 and F4) output by the second and third convolutional blocks to obtain a low-level feature map F low and a high-level feature map F high Input into the cross-fusion block, and use bilinear interpolation operation to change the resolution of the high-level feature map F high to the same resolution as the low-level feature map F low Refine and upsample the high-level feature map through a 3×3 convolutional layer with BN and ReLU; concatenate the upsampled high-level feature map F high and the low-level feature map F low Perform concatenation, and generate the output of the cross-fusion block through a 1×1 convolutional layer with BN and ReLU.

7. A rasterized point cloud pixel-level classification system based on deep learning according to claim 5, characterized in that: The feature enhancement unit includes an object area feature generation subunit and a calculation aggregation subunit. The object region feature generation subunit: is used to divide the feature map obtained through information refinement and fusion into k object regions {M1, M2, …, M k}, aggregate all pixels of each object region, and weight them according to the degree of their belonging to the object region, to obtain the object region feature f k ; The calculation and aggregation subunit: is used to calculate the relationship ω between each pixel and each object region ik , by aggregating the representations of k object regions and through the object region feature f k and the relationship ω between the pixel and the region ik , the enhancement of the feature is realized through a 1×1 convolutional layer with BN and ReLU, and finally the features of the original representation and the object context representation after feature enhancement are obtained.

Images